Multi-model motion tracking under multiple team member actuators

Autonomous robot agents need to be able to track moving objects. When tracking is performed by a robot executing specific tasks acting over the target being tracked, such as a Segway RMP soccer robot grabbing and kicking a ball, the motion model of the target becomes dependent on the robot's actions. The robot's tactic provides valuable information in terms of the target behavior. We introduced a play-based motion modeling and tracking in such scenarios (Gu & Veloso, 2006). Observations from the sensors might consist of multiple measurements due to the moving objects and the clutter. Generally the clutter has similar color as the targets we are interested in and it causes multiple hypotheses for the true targets. Our previous approach does not perform well once incorrect measurements originating from clutter or false alarms exist, which causes multiple hypothesis of the tracked target. Recently, a hy brid approach for online joint detection and tracking for multiple targets was proposed (Ng et al., 2005). This approach does not assume the knowledge of true targets (without clutter) is given. It first uses a deterministic clustering method that searches for regions of interest (ROIs) based on the observations and monitors these ROIs for target detection, then performs multi-target tracking by Sequential Monte Carlo (SMC) methods. In this chapter, we extend the contributed play-based tracking by introducing a multi-target dynamics model. We also take use of an improved proposal function for the PBPF based on the ROIs. We construct two multi-target trackers in the system, for the ball and the team member respectively. We use multi-target tracker instead of single target tracker because we can keep track of the true target and the false positive at the same time without losing any of them and perform ball (or team member) recognition from a pool of tracked objects later. This chapter is organized as follows. We give a brief description of our robots and two main components of the control architecture. We describe the multi-target dynamics model. We describe the clustering algorithm we used to continuously monitor the appearance and disappearance of regions of interest (ROIs) on the field. The ROIs is used to deterministically obtain the number of targets. The ROIs is fu rther used to get better proposal functions in particle filtering algorithm. We use an improv ed proposal function for the PBPF based on the ROIs. We contribute the multi-target tracking extension to the PBPF introduced in (Gu & Veloso, 2006).